Prediction of Protein-Protein Interaction Sites Using Constructive Neural Network Ensemble

Abstract

Prediction of protein-proteininteraction sites is very important to the function of a protein and drug design. In this paper, we adequately utilize the characters of ensemble learning, which can improve the accuracy of individual classifier and generalization ability of the system, and propose a new prediction method of protein-protein interaction sites: ensemble learning method based on the constructive neural network. Protein sequence profile and residue accessible area are used as input feature vectors. We evaluate the ensemble classifiers and compare them with several traditional methods (SVM, ANN, CA and Bayesian) on the dataset of 61 protein chains with 5-fold cross validation. The results clearly show that the proposed ensemble method is quite effective in predicting protein binding sites. Our method achieves good performance (Accuracy of 73.26%, Sensitivity of 58.38%, Specificity of 68.87%, CC of 35.47% and F1-measure of 63.04%), which is significantly better than that of the compared methods. The results obtained show that our proposed method is a promising approach for predicting protein-protein interaction sites.The experiments show the validation and correctness of the ensemble method based on Covering Algorithm (CA).